Showing 1–1 of 1 results for author: Bergh, P
-
Experiences with Remote Examination Formats in Light of GPT-4
Authors:
Felix Dobslaw,
Peter Bergh
Abstract:
Sudden access to the rapidly improving large language model GPT by open-ai forces educational institutions worldwide to revisit their exam procedures. In the pre-GPT era, we successfully applied oral and open-book home exams for two courses in the third year of our predominantly remote Software Engineering BSc program. We ask in this paper whether our current open-book exams are still viable or wh…
▽ More
Sudden access to the rapidly improving large language model GPT by open-ai forces educational institutions worldwide to revisit their exam procedures. In the pre-GPT era, we successfully applied oral and open-book home exams for two courses in the third year of our predominantly remote Software Engineering BSc program. We ask in this paper whether our current open-book exams are still viable or whether a move back to a legally compliant but less scalable oral exam is the only workable alternative. We further compare work-effort estimates between oral and open-book exams and report on differences in throughput and grade distribution over eight years to better understand the impact of examination format on the outcome. Examining GPT v4 on the most recent open-book exams showed that our current Artificial Intelligence and Reactive Programming exams are not GPT v4 proof. Three potential weaknesses of GPT are outlined. We also found that grade distributions have largely been unaffected by the examination format, opening up for a move to oral examinations only if needed. Throughput was higher for open-book exam course instances (73% vs 64%), while fail rates were too (12% vs 7%), with teacher workload increasing even for smaller classes. We also report on our experience regarding effort. Oral examinations are efficient for smaller groups but come with caveats regarding intensity and stress.
△ Less
Submitted 27 March, 2023;
originally announced May 2023.
